Joined: 14 Jan 2005 Posts: 1494 Location: Delphi, Indiana
Posted: Thu Sep 27, 2007 10:41 am Post subject:
asimpkins wrote:
Quote:
...It seems to be a better proxy of being big and near the basket...
You're right, DReb are mostly available to big guys near the basket. In that sense, the guards may not get their share of 'defensive credit from DReb'. Then again, a player like Kidd (not that there's anyone 'like' Kidd) should be distinguished from a non-rebounder/defender, and his rebounding helps to do this.
Since purely defensive stats are few and far between, DReb are ripe for being the default indicator. Can we find examples of 'absurd results' that overamplify someone's contributions by the # of DReb they get?
Meanwhile, players who aren't going for rebounds should be out on the break, or otherwise leading the offensive. This gives them first crack at scoring/assisting. In general a player is either a ballhandler or a rebounder, and in either case they have the earlier opportunity to get on the statistical scoreboard (either on O or on D).
Gross generalizations rule! _________________ 40% of all statistics are wrong.
Defensive rebounding is overcreditted when meaningful defensive pressure has been applied resulting in not only a miss but a expected FG% below average or replacement level.
But you could argue (not that Berri's or PER or anysystem lacking shot defense have) that a wide open shot or token shot pressure may not be that impactful while garnering the possession certainly is. Ed Petersen/s article at 82 games scored about 60% of all Kings shots as open or wide open. http://www.82games.com/saccon.htm
Kings scoring efficiency was roughly 50% -100% higher when open or wide open when compared to contested or heavily contested. Kings scoring efficiency allowed was roughly twice high when shot defense was light or non-existent. If you had leaguewide averages maybe you could develop a simple 2 part credit split approximation among just actual shot defender(s) and the rebounder based on whether the shot was adequately contested and that leaguewide average efficiency impact, giving the defender his stat based contribution and giving the rebounder the remainder.
The only operating split credit/blame systems generally split all credit and blame equally except in some cases with assists. Rule based on and off the ball variable/partial credit/blame systems could be implemented (not perfect but an attempt to get closer to what happens and what really matters) but have been hard to garner agreement about or support for.
Teams can work video hard. I don't know how they score these, if it generally gets reduced to did each player do their job well enough as a whole or not or a yes/no scoring of specific component part of defense (movement, hands, rotations, closeouts, etc.) but an alternative would be score each possession say on a 10 point credit/blame system. Perhaps typically everyone getting at least 1 credit (splitting at least half the 10 points) but you'd wouldnt have to be rigid about that. You could vary the shooter/assist ratio and give credits to the secondary passer, pick setter, spacer, post threat, etc as appropriate to the eye of a knowledgeable scorer. Split credit, but subjective since rules can't perfectly capture everything. But of course consistent "evaluation standards" would still have some appeal and could be difficult to achieve beyond one team / one videotape scoring operation.
Many super intensive advancement of knowledge paths face proprietary interests that would shield the results for competitive gain. I take it baseball fanatics are doing a better job of supporting publc domain, intensive/advanced stats, at least for now. Synergy offered the possibility of a great leap forward but my impression is it remains in hands of insiders and hasn't realized the absolute fullest merger of video and stats possible (particularly on defensive side of the ball). All that is possible may not become public but perhaps more will develop.
Last edited by Mountain on Thu Sep 27, 2007 1:37 pm; edited 6 times in total
Can we find examples of 'absurd results' that overamplify someone's contributions by the # of DReb they get?
I'm out of my element when talking about a player's non-statistical defensive contributions. I don't watch most players enough to have a valid opinion, so I can really only go on reputation. Looking through this list of rebounders... is Zach Randolf one of the 15th best intangible help defenders? Is Al Jefferson top 5?
And, of course, players like Bowen (160th) get really screwed.
Mountain wrote:
Defensive rebounding is overcreditted when meaningful defensive pressure has been applied resulting in not only a miss but a expected FG% below average or replacement level. But you could argue... that a wide open shot or token shot pressure may not be that impactful while garnering the possession certainly is.
Even if the defense has been token that doesn't mean that the rebounding hasn't been equally token. How often does the rebound just simply randomly bounce to a uncontested player who needs to do nothing more than catch it -- perhaps while his teammates fight to box out the opposition? Not all boards are competitive tests of strength, hustle, timing, and athleticism.
Box outs do have value and deserve mention and ideally credit share.
A defensive possession that ends in a miss and a defensive rebound under a 10 point multi-share scoring system might give on a particular play +1 of 10 to the defender of each of 3 players who made any sort of scoring threat, maybe +2 more to the actual defender of the shot that missed, maybe +3 to the actual rebounder and maybe +1 each to the 2 guys that boxed out their offensive man and allowed the actual rebounder a good shot at winning the ball. There is room for any of a range of combinations of defensive credit spilts. If a simple 10 share is too tight to divide finely enough you could make it a 100 point system. Conceptually.
Last edited by Mountain on Mon Oct 01, 2007 3:12 pm; edited 1 time in total
Joined: 14 Jan 2005 Posts: 1494 Location: Delphi, Indiana
Posted: Fri Sep 28, 2007 8:04 am Post subject:
asimpkins wrote:
.. is Zach Randolf one of the 15th best intangible help defenders?..
.
Well, Zach's RbR went up 26% between '06 and '07, and the Blazers coincidentally went from dead last in DReb% to just-average. He didn't just work on his shot, but also his boardwork; the team improved on the scoreboard, from -9.5 PPG to -4.2 PPG.
Securing the DReb is perhaps (??) half of the team's defensive job. About 30% of missed FG are OReb, which have a very high Pts/Poss rate. Per-event, one could argue the DReb is at least as important as forcing the low-% shot.
So (I'm thinking) only some fraction of Zach's DReb value can be attributed to 'help defense'. Most of it is just rebounding. A few might be 'easy' rebounds, but a few baskets are 'garbage', or assists, or steals. _________________ 40% of all statistics are wrong.
Joined: 13 Oct 2005 Posts: 359 Location: Atlanta, GA
Posted: Fri Sep 28, 2007 8:54 am Post subject:
Mike G wrote:
asimpkins wrote:
.. is Zach Randolf one of the 15th best intangible help defenders?..
.
Well, Zach's RbR went up 26% between '06 and '07, and the Blazers coincidentally went from dead last in DReb% to just-average. He didn't just work on his shot, but also his boardwork; the team improved on the scoreboard, from -9.5 PPG to -4.2 PPG.
But their offense was the chief reason for the team's improvement; they shot better, committed fewer turnovers, got to the line and made FTs more often, and, yes, rebounded the ball much better on O.
Their defense barely improved, though. If Zach's rebounding improvement at that end was such a big deal, why did the team still rank 26th in defensive efficiency (while ranking 15th in defensive rebounding) a year after ranking 29th in defensive efficiency (while ranking 30th in defensive rebounding)? And why was their defense worse with him on the floor than off, again? His defensive rebounding rate went from 19.1% to 25.6%, and yet his defensive +/- impact on the Blazers got worse, +1.5 to +1.7.
Joined: 03 Jan 2005 Posts: 497 Location: Greensboro, North Carolina
Posted: Mon Oct 01, 2007 6:53 am Post subject:
Here is the promised PowerPoint presentation, although it is not really a substitute for the yet unfinished paper. And it is less detailed that my posts on this topic have been.
I thought the presentation went OK both here and when I gave it as a brown bag at the Council of Economic Advisers, but it still reflects the fact that it is very much a working paper.
Probably the biggest new finding here is that the one-year Adjusted Plus/Minus metric (roughly the metric that Winston and Sagarin use) is by far the worst metric here in terms of predicting future wins or future adjusted plus/minus, even worse than Wins Produced (which is the second worst). But the noisiness of the one-year adjusted plus/minus metric is something I have harped on since the first time I started posting on plus/minus back at the old board.
Sorry I don't have much more to say here, but I just don't have the time to parse every word to the level of detail needed, given that Henry at TrueHoop has a tendency to use my posts here to escalate the conflict between Berri and me - something I have tried to de-emphasize when I have presented the paper.
Joined: 14 Jan 2005 Posts: 1494 Location: Delphi, Indiana
Posted: Mon Oct 01, 2007 7:42 am Post subject:
Berri has apparently made a major effort to get his views repeated by Henry.
Henry cites Berri as if he were an acknowledged authority. In fact, I've yet to read a single corroborator who both (a) understands, and (b) believes that Berri is saying anything.
The (few) Berri supporters here (at APBRMetrics) invariably plead ignorance when confronted with an inconsistency or implausibility. Berri doesn't answer direct lines of questioning such as those asked consistently by Dan. In other words, Berri is only supported by himself and by people who don't understand his methodology.
This would not seem to bode well for a theory that's been discussed for over a year. I don't think there's a split in the metrics community; but that Berri's attempt is a branch that has failed to graft. _________________ 40% of all statistics are wrong.
Can't see the powerpoint slide, not having powerpoint on computer. Any workarounds? Look forward to any comments from those in attendance and eventually seeing the paper.
Dan's comment about pure adjusted +/-'s low predictive power adds further reason for me to de-emphasize casual use of it, at least some, in analyzing player impact. I've always had reservations given the simple input and 10 players contrbuting to that input data regardless of the power of method afterwards.
(I'd like to try to understand statistical +/- and overall adjusted +/- better and in context with other regression based findings. I could try to understand Eli's spreadsheet formulas better and where they came from. But I also came back across this note from a year ago by Dan on his statistical plus/minus metric: "Now I will not go into detail about how it is computed, but it does incorporate complicated adjustments for position and for other roles besides position" So full understanding of statistical +/- or even overall +/- is unlikely to be acquired independently. Oh well.)
Last edited by Mountain on Mon Oct 01, 2007 7:47 pm; edited 4 times in total
Joined: 03 Jan 2005 Posts: 660 Location: Washington, DC
Posted: Mon Oct 01, 2007 1:11 pm Post subject:
Dan: I could use more explanation for everything, but I'm particularly interested in Table 2. In the "Predicting" section, the first column seems straightforward -- basically it's testing the predicting power of each metric -- its ability to predict future wins. The 2nd column is whether that metric can predict +/-.
Why is predicting adjusted +/- a more powerful test than predicting team wins? Why would we be interested in predicting future +/- when it doesn't do a very good job of explaining wins?
Why does adjusted +/- do a worse job than any of the other metrics at predicting future +/-?
As far as a Power Point work-around -- simplest thing might be to make it into a pdf and host it somewhere. _________________ My blog
Thanks very much for the suggestion Eli. My OS was too old for that version but I downloaded the older one and it worked.
Following on kjb comments it looks like with the exception of "pure" adjusted +/- the metrics included in the study are tightly bunched in explaining wins which is often the main purpose. Alt Wins Produced is the slight winner on that basis over the simple metric of minutes.
Is it correct that Alt Wins Produced is still consistent with "WS2PT = Win Score with a position and team adjustment, where the weight on field goals missed has been changed to -0.7, the weight on free throws missed has been changed to -0.35, the weight on offensive rebounds has been changed to 0.7, and the weight on defensive rebounds has been changed to 0.3."?
Explaining +/- could be important to bettors or I guess it could be important in thinking about playoffs though that opens other issues about translation.
Minutes was 2nd worst at explaining +/- on Table 2 in stark contrast to leading performance on table 1 and column 1 of table 2. Not sure what to make of that.
(Scores of statistical +/- and overall +/- as well as other known or experimental metrics in relation to the tested left for speculation or future study as expected.)
Joined: 03 Jan 2005 Posts: 497 Location: Greensboro, North Carolina
Posted: Tue Oct 02, 2007 10:34 pm Post subject:
kjb wrote:
Dan: I could use more explanation for everything, but I'm particularly interested in Table 2. In the "Predicting" section, the first column seems straightforward -- basically it's testing the predicting power of each metric -- its ability to predict future wins. The 2nd column is whether that metric can predict +/-.
Why is predicting adjusted +/- a more powerful test than predicting team wins?
It is simply an empirical matter. There are more individuals than teams to predict, even with fewer years of +/- data. And there is more variation among players than there is among teams. The additional observations and variation result in a more powerful test.
Quote:
Why would we be interested in predicting future +/- when it doesn't do a very good job of explaining wins?
Past wins do a poor job predicting future wins and are worse than +/- at predicting future +/-. How good a metric is at being a predictor doesn't tell us much about whether that metric is useful as an outcome.
As an outcome, +/- uses hundreds of players over multiple seasons. As a predictor, +/- uses one player from one season. In those two cases there is a huge difference in the consequences of the low signal-to-noise ratio of +/-. And remember that the signal here picks up pretty much all the contributions a player makes while he is in the game and the noise is unrelated to any of the other metrics.
So the folks who claim that using +/- as an outcome is nothing more than examining "how well each model predict[s] [my] model" just don't get it. They are, in essence, saying that how the scoreboard changes while a player is in the game is irrelevant. I think that is a peculiar claim that has nothing to do with whether or not +/- is a good way to evaluate players.
Quote:
Why does adjusted +/- do a worse job than any of the other metrics at predicting future +/-?
It is the same story I have told countless times. One-year adjusted plus/minus is really, really noisy. I won't go into detail here, but I have suggested numerous ways of dealing with the noise.
The ability to import some new values into Wins Produced and improve the performance of Alt Wins Produced to best in show of this group demonstrates at a general level the potential value of hybrid approaches.
With a longterm dataset and the time & desire an endless range of metrics and meta-metrics could be tested and better and better fit / performance could be discovered.
Every new season will be somewhat different but there is plenty to justify NBA team investment in analytical research.
(Of course the games have to be played and will be unique based on real-time decisionmaking, skill and chance.) Further analytic quest by fans? That's up to each.
Last edited by Mountain on Wed Oct 03, 2007 12:44 pm; edited 1 time in total
The ability to import some new values into Wins Produced and improve the performance of Alt Wins Produced to best in show of this group demonstrates at a general level the potential value of hybrid approaches.
Just playing off of this comment for a second.
I'm not sure that it makes sense to call that a "hybrid" approach. I think that term should be reserved for systems that use multiple different types of data (like DCS and OCS earlier this summer).
Part of what became very clear to me working with linear weight systems is that they are, by definition, a series of linear weights. All of the discussion is about the ways to derive/justify the weights but, essentially one system is the same as the next except for the weights.
To contradict myself, Wins Produced isn't strictly a linear weights system. The position adjustment, team adjustment, and assist adjustment are all separate from the base linear weights.
I just think it's important to be clear about what is actually being done in the various systems.
You cannot post new topics in this forum You cannot reply to topics in this forum You cannot edit your posts in this forum You cannot delete your posts in this forum You cannot vote in polls in this forum
FAQ
Search
Memberlist
Usergroups
Register
Profile
Log in to check your private messages
Log in
